Scholarlink

<p>Real-time speech communication systems, including telemedicine platforms, online conferencing tools, and smart voice interfaces, are highly susceptible to non-stationary background noise, which degrades intelligibility and perceptual quality. While recent deep learning models achieve strong enhancement performance, many are computationally intensive and unsuitable for latency-constrained deployment. This work investigates whether a lightweight feed-forward deep neural network (DNN), augmented with noise-aware training and global variance (GV) equalization, can achieve a favorable balance between enhancement quality and real-time feasibility. The proposed framework performs spectral-domain log-magnitude mapping using a noise-aware conditioned multilayer perceptron, with mean noise statistics appended to the input features to improve robustness under non-stationary noise. A global variance equalization step is applied as post-processing to mitigate spectral over-smoothing commonly observed in regression-based models. The system is trained on a synthetically generated dataset comprising approximately 2500&#xa0;h of noisy-clean speech mixtures derived from a compact clean speech subset combined with diverse noise types and signal-to-noise ratios. The system is evaluated against classical and lightweight neural baselines using objective metrics, including PESQ, STOI, segmental SNR, and log-spectral distance, as well as a controlled subjective Mean Opinion Score (MOS) study. Results demonstrate consistent improvements over baseline methods while maintaining low-latency inference compatible with real-time processing, achieving real-time factors of 0.54 on GPU and 0.98 on CPU. These findings indicate that efficiency-driven, noise-aware spectral enhancement can provide a practical solution for real-time communication scenarios where deterministic latency and hardware feasibility are primary constraints.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Real-Time Speech Denoising in Communication System Using Deep Learning Under Non-Stationary

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • Samuel Yaw Mensah,
  • Tao Zhang,
  • Yangzhang Geng,
  • Nahid AL Mahmud,
  • Stephen Afrifa

摘要

Real-time speech communication systems, including telemedicine platforms, online conferencing tools, and smart voice interfaces, are highly susceptible to non-stationary background noise, which degrades intelligibility and perceptual quality. While recent deep learning models achieve strong enhancement performance, many are computationally intensive and unsuitable for latency-constrained deployment. This work investigates whether a lightweight feed-forward deep neural network (DNN), augmented with noise-aware training and global variance (GV) equalization, can achieve a favorable balance between enhancement quality and real-time feasibility. The proposed framework performs spectral-domain log-magnitude mapping using a noise-aware conditioned multilayer perceptron, with mean noise statistics appended to the input features to improve robustness under non-stationary noise. A global variance equalization step is applied as post-processing to mitigate spectral over-smoothing commonly observed in regression-based models. The system is trained on a synthetically generated dataset comprising approximately 2500 h of noisy-clean speech mixtures derived from a compact clean speech subset combined with diverse noise types and signal-to-noise ratios. The system is evaluated against classical and lightweight neural baselines using objective metrics, including PESQ, STOI, segmental SNR, and log-spectral distance, as well as a controlled subjective Mean Opinion Score (MOS) study. Results demonstrate consistent improvements over baseline methods while maintaining low-latency inference compatible with real-time processing, achieving real-time factors of 0.54 on GPU and 0.98 on CPU. These findings indicate that efficiency-driven, noise-aware spectral enhancement can provide a practical solution for real-time communication scenarios where deterministic latency and hardware feasibility are primary constraints.